jing 2008
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
tent and the
key factor for
it when image
of gradient is
; grey-level is
i>bert, Prewitt,
or calculating
bel arithmetic,
(8)
Hows:
(9)
dded together,
hole image is
the change of
; gradient can
experiment is
:ions of image
tion matching
lation between
al information
rom 4.0 e+005
»tained, shown
magnitude has
no strong relation to the success-matching rate for mutual
information based method. Even if there are much less features in
the image, good matching result can also be obtained using mutual
information matching approach.
Table 2. The relationship between image gradient and matching
success rate
Image
gradient
Number of
input images
Number of
correct matching
Success rate
<%)
4.0—4.2
20
11
55.00
4.2—4.4
27
14
51.85
4.4—4.6
34
25
73.53
OO
1
33
22
66.67
O
1
OO
30
23
76.67
5.0—5.2
31
22
70.97
5.2—5.4
19
17
89.47
5.4—5.6
28
26
92.86
Os
1
bo
26
24
66.67
5.8—6.0
19
16
84.21
4.3 Self-Similar Pattern
In image matching, self-similar pattern in reference image
seriously affects success rate of matching. Self-similar pattern
often indicates some sub-area, which grey or some features appear
in the reference image repeatedly. Different method has different
definition about self-similar pattern. For area-based method, the
definition of self-similar pattern is the number of sub-areas, which
have similar grey level distribution. Whereas, for feature-based
methods, the definition of self-similar pattern is the number of sub-
areas, which have similar feature distribution (Xie, et al., 1997).
Suppose a sub-image i is selected from reference image, its self
similar pattern is defined as follows:
cfi=^~ (10)
Where 5 is the number of all sub-images, which are used for
matching when sub-image i is searching on the reference image
pixel by pixel. Suppose the size of a reference image is MxN
pixels, and the size of a sub-image is mxn pixels, then,
s = (M - m +1) x (TV - n +1), where Pj is the number of sub-images
in 5 , whose grey correlation coefficient obtained by matching
them to sub-image i is greater than the threshold TH .
It is known that c/ ( denotes two dimensions relativity about a sub-
area to the whole area. If a number of sub-images are cropped
equably from reference image, then, the mean self-similar pattern
value of these sub-images can be used to inspect two dimensions
relativity of reference image. Therefore, self-similar pattern value
of a reference image is defined as the following equation:
i=1
Where / is the number of sub-images cropped from the reference
image. And these sub-images must satisfy some requirements.
The bigger the self-similar pattern value is, the stronger two
dimensions relativity of image self is, and the higher the error of
matching is. In the calculating of self-similar pattern value, TH
and n are two important parameters. Selection of TH is a key
problem. Only when a reasonable TH is selected, characteristics
of image’s self-matching can behave fully. The Selection of n
affects the calculation time and reliability of self-similar pattern
value. Sampling interval determines how much n is. Setting
TH = 0.96 and n = 20 are made after experiments in the paper.
The relationship between self-similar pattern and success rate of
matching based on mutual information is shown in Table 3 and
Figure 5.
Table 3. The relationship between self-similar pattern value for
different scene and success rate of matching
Different scene
image
Self-similar pattern
value
Success rate of
matching (%)
Farmland
0.989
69
Village
0.519
78
River
0.336
81
Town
0.117
93
Stream
0.010
100
Self-similar pattem value
Figure 5. Self-similar pattern and success rate of image matching
From Figure 5 we can see that the bigger the value of self-similar
pattern is, the lower the success rate of matching is, since there are
many self-similar areas in the reference image, which leads to
many mistakes in matching and reduce the validity of matching
based on MI. Self-similar pattern means how many self-similar
areas are in one image, the more the self-similar areas there are,
the more the peaks of MI value there are, and local maximum of
MI will cause matching error. There is a strong relativity between
matching success rate and self-similar pattern.
4.4 The Validity of Mutual Information Similarity Metric
The validity of mutual information similarity metric applied in
template matching for dissimilar image can be validated by success
rate of matching. The success rate of matching shows the ability of
1133